Harness

The building block for driving a coding agent from inside your loop. app.harness(prompt, schema=...) runs CodeAF, AgentField's native harness, by default -- or Claude Code, Codex, Gemini CLI, OpenCode, Pi, or OMP when you choose them. Either way it waits for structured output, validates it against your schema, and hands back a rich result with cost and retry metadata.

CodeAF can be provisioned by the installer, Desktop, or the SDK on first execution. Set OPENROUTER_API_KEY; use af codeaf ensure to pre-install before the call below. See provisioning and upgrades.

Without harness, you would manage subprocess plumbing, JSON repair, retry-with-backoff, schema validation, output-file cleanup, and per-provider quirks yourself. With harness, those become one call you can wrap in a loop, a tournament, or an adversarial pattern.

The shape

from pydantic import BaseModel
from agentfield import Agent, HarnessConfig

class ReviewResult(BaseModel):
    findings: list[str]
    severity: str  # "low" | "medium" | "high"

# Defaults pinned on the agent — every .harness(...) call inherits them.
# No provider, no model: this runs on CodeAF with `CODEAF_MODEL`, legacy `AFORGE_MODEL`, or the built-in default model.
app = Agent(
    node_id="reviewer",
    harness_config=HarnessConfig(
        max_turns=12,
    ),
)

@app.reasoner()
async def review_diff(diff: str) -> dict:
    result = await app.harness(
        f"Review this diff. Be precise, no fluff.\n\n{diff}",
        schema=ReviewResult,
    )

    if result.is_error:
        return {"ok": False, "error": result.error_message}

    return {
        "ok": True,
        "review": result.parsed.model_dump(),
        "cost_usd": result.cost_usd,
        "num_turns": result.num_turns,
        "session_id": result.session_id,
    }

Choose your worker

The default is a starting point, not a lock-in. Set provider and the same call runs on a different coding agent — that is how you orchestrate Claude Code, Codex, Gemini CLI, or OpenCode from inside an AgentField loop, or run a fleet of them side by side.

result = await app.harness(prompt, schema=ReviewResult, provider="claude-code")

# Claude Code is also the provider that enforces a USD cost cap and a tool allowlist.
capped = await app.harness(
    prompt,
    schema=ReviewResult,
    provider="claude-code",
    permission_mode="plan",
    max_budget_usd=0.50,
    tools=["Read", "Grep", "Glob"],
)

CodeAF enforces turn and time limits only. It ignores max_budget_usd, tools, and permission_mode; use max_turns and AGENTFIELD_HARNESS_TIMEOUT_SECONDS (default 1800).

Each override needs its own worker installed and its own credential. CodeAF needs OPENROUTER_API_KEY and either a pre-installed binary or first-call download.

Provider selection resolves in this order, first match wins:

1. An explicit provider on the call, then on the agent's HarnessConfig.
2. The AGENTFIELD_HARNESS_PROVIDER environment variable -- codeaf, aforge (alias), claude-code, codex, gemini, opencode, pi, or omp (the Python SDK also accepts grok).
3. codeaf.

HarnessConfig — defaults on the Agent

Pin defaults on the agent constructor; override per call only for the dimensions that genuinely vary by task.

HarnessResult — what comes back

Schema-bound output

Pass a Pydantic model (Python), Zod schema (TypeScript), or Go struct as schema=. The runner injects an OUTPUT REQUIREMENTS suffix telling the agent to write JSON to .agentfield_output.json, then reads, repairs, validates, and returns result.parsed as a validated instance. Three recovery layers run before declaring failure:

1. Parse the output file directly.
2. Cosmetic repair — strip markdown fences, trailing commas, repair truncated braces.
3. One-shot AI repair — re-emit the same content as valid JSON conforming to the schema (no tools, no exploration; cheap reformatting only).

After that, the run is retried up to max_retries times. The output file is cleaned up automatically.

The agent writes its answer to a file with its own Write tool instead of the harness parsing JSON out of response text. Every provider has a Write tool, so the same strategy works on all of them:

Provider switching is a one-field flip

Every provider is interchangeable through provider=. Same loop code, different worker:

# Plan with Claude (careful reasoner), execute with Codex (fast implementer).
plan = await app.harness(prompt, provider="claude-code", permission_mode="plan", schema=ChangePlan)
edits = await app.harness(apply, provider="codex",       permission_mode="auto", schema=EditReport)

See the provider docs for the role each one plays best:

• CodeAF: the default. Pre-install with af codeaf ensure or allow a first-call download.
• Claude Code: the careful reasoner.
• Codex: the fast implementer.
• Gemini CLI: the long-context worker.
• OpenCode: the open-model path.

Error handling

A failed run does not raise. Check is_error, then branch on failure_type to tell the failure modes apart:

result = await app.harness("Deploy the staging environment.", max_turns=20)

if result.is_error:
    match result.failure_type:
        case "timeout":
            log.warning(f"Harness timed out after {result.duration_ms}ms")
        case "crash":
            log.error(f"Agent crashed: {result.error_message}")
        case "schema":
            log.warning("Output did not match schema after retries")
        case "api_error":
            log.error("Transient API error, will retry on next invocation")
else:
    print(result.text)

In Go, compare result.FailureType with harness.FailureTimeout, harness.FailureCrash and harness.FailureSchema.

Common patterns

Retry with provider fallback

When the primary provider trips a budget or rate-limit, fall back to another worker without losing the loop's intent. Starting from the default keeps the happy path install-free.

async def hardened_harness(prompt: str, schema, **opts):
    for provider in ("codeaf", "claude-code"):
        result = await app.harness(prompt, schema=schema, provider=provider, **opts)
        if not result.is_error and result.parsed is not None:
            return result
    raise RuntimeError(f"all providers failed: {result.error_message}")

Tournament

Run multiple providers in parallel against the same prompt and pick the best result with an LLM-as-judge call.

runs = await asyncio.gather(*[
    app.harness(prompt, provider=p, schema=ImplResult, max_budget_usd=0.05)
    for p in ("claude-code", "codex", "opencode")
])

verdict = await app.harness(
    "Pick the best implementation. Score on correctness and minimality.\n\n"
    + format_runs(runs),
    provider="claude-code",
    schema=Verdict,
)
return runs[verdict.parsed.winner_index]

Adversarial verifier

One provider proposes a change, a second one verifies it. The verifier's output triggers re-runs until severity drops or the budget caps out.

for attempt in range(3):
    impl = await app.harness(prompt, provider="codex", schema=ImplResult)
    if impl.is_error:
        continue

    review = await app.harness(
        f"Find real bugs in this change:\n\n{impl.parsed.diff}",
        provider="claude-code",
        permission_mode="plan",
        schema=ReviewResult,
    )
    if review.parsed.severity in ("low", "medium"):
        return impl  # accepted

    prompt = f"Re-attempt. Previous reviewer findings:\n{review.parsed.findings}\n\n{prompt}"

Authentication

Each provider reads its own credentials from the environment forwarded to the harness subprocess. None of these need to be passed through harness_config.env if they're already in the process environment.

Check what a provider can actually see before you run a loop against it:

af harness doctor --provider codeaf     # or claude-code / codex / gemini / opencode

See also

• Cross-agent calls — call other AgentField nodes from inside a harness loop, or vice versa.
• CodeAF: the default harness
• Claude Code
• Codex
• Gemini CLI
• OpenCode